co-evolving trust mechanisms for catering user...
TRANSCRIPT
adfa, p. 1, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Co-evolving trust mechanisms for catering user behavior
Tanja Ažderska
Laboratory for Open Systems and Networks, Jozef Stefan Institute, Jamova 39,
1000 Ljubljana, Slovenia
Abstract. While most of the computational trust models devote to truthfully de-
tecting trustworthy individuals, much less attention is paid to how these models
are perceived by users, who are the core of the trust machinery. Understanding
the relation between trust models and users‟ perception of those models may
contribute for reducing their complexity, while improving the user-experience
and the system performance. Our work recognizes reputation, recommendation
and rating systems as online trust representatives and explores the biased beha-
vior resulting from users‟ perception of those systems. Moreover, we investi-
gate the relation and inter-dependencies between trust mechanisms and user be-
havior with respect to context, risk, dynamics and privacy. We perform experi-
mental study and identify few types of cognitive biases that users exhibit. Based
on the identified factors and the findings of the study, we propose a framework
for addressing some of the issues attributed to users‟ biased behavior.
Keywords: trust, bias, context, reputation, recommendation, rating
1 Introduction
Few decades ago, trust was a feeling and a reality. The curiosity of „feelings often
diverging from reality‟ made trust a major constituent of social studies. Person A may
believe that B is trustworthy, although that is not the case, but A may also believe in
the opposite (B being non-trustworthy), although it may not be the case. This led to
the creation of social models of trust. Nowadays, trust has also come to represent
people‟s beliefs in the entities met in the virtual world, leading to the design of many
computational models of trust. However, the models that represent the relationship
between A and B can easily fail to capture both sides of the story: how much A trusts
B, and whether B is really trustworthy. In other words, how much the model resem-
bles reality, how much it comes closer to human perceptions and actions, and how
much the two differ. In this study, we take the stance that trust has its own representa-
tives in the online environment. We recognize reputation, rating, and recommender
systems (henceforth denoted as RRR systems) as the online representative of trust.
We see the elements of the three (RRR) systems as ones that are different, but in often
complement one another. However, the adequate combination of RRR depends on
many factors, and the failure to recognize those factors leads to inconsistencies in the
work of the RRR system as a whole. Moreover, user behavior is largely influenced by
the work of those systems, and it largely influences the systems performance as well.
The contribution of our work is in the following: we detect and analyze four factors
that influence the work of RRR systems when they are required to co-evolve as a
single solution for providing good user experience, system reliability and result accu-
racy. Those factors are context, risk, dynamics and privacy. We put users in the core
of RRR systems, and investigate the inter-dependencies between RRR systems and
user behavior with respect to the four identified factors. We analyze several types of
cognitive biases that users exhibit in their online experiences with RRR systems.
Then, we perform an experimental study that will allow us to identify several types of
cognitive biases exhibited by the users, and that were not investigated in an RRR
setting so far. Based on the identified factors and the findings of the experimental
study, we propose a framework for addressing some of the issues attributed to users‟
biased behavior and explore the possibility of employing hidden signals in the RRR
systems for the purpose of capturing some types of user behavior.
To provide a clear picture of our understating of social trust and its online repre-
sentatives, in the following section we define the concepts. Then, we present related
work in the area with the effort to move closer to each other the social and the tech-
nical aspect. Finally, we present the stated contributions throughout Sections 4 and 5.
We conclude in Section 6.
2 The Notion of Trust and its Online Representatives
2.1 Trust
From a social perspective, trust can be defined from two general aspects: cognitive
and affective. The former is represented by concepts like rational choices, learning
loops, institutional protocols, pattern detection, and imitation of established norms.
Affective aspects, on the other hand, are mainly seen in the emotional side of trust
interactions, and they account for the human feelings. As feelings are heavier on
energy, whereas thoughts are heavier in information, affective properties often “take
the blame” for contributing to cognitively biased decisions [1]. The „social‟ literature
on trust and reputation is exhaustive [1–3]. Clearly, when seeing trust as a purely
social phenomenon, it can only be ascribed to living beings, and manifested through
the property of trustworthiness of entities that are not necessarily living beings them-
selves. However, the ability to trust nowadays is disentangled from purely social con-
texts. A trusting entity can be any agent capable to resolve cognitive conflicts, or do
preferential filtering, ranking, or sorting. This makes the definition and even the pur-
pose of trust hard to grasp and determine. The blurred line of where the human factor
starts or stops to influence trust and trustworthiness, usually leads to neglecting the
affective side when defining the computational analog of trust. This, in turn, leads to
the inability to predict the behavior of systems where trust models are deployed.
These effects are caused by the highly non-linear nature of the trust phenomena,
which do not allow a system to be designed according to the elegant principles of
mathematical linearity and probabilistic averaging. Hence, knowing the composition
of the system parts does not contribute a lot to inferring the properties of the system
as a whole. It is critical to consider the interactions and dependencies between the
entities that comprise the system, and capture the additional phenomena and proper-
ties that emerge from those interactions. Complementing this with the strong contex-
tual dependence of trust explains why researchers have a hard time formalizing trust,
and incorporating it into online scenarios analogous to those in the traditional social
networks. Following Gambetta [4], we give the following initial definition:
Definition 1a. Trust is the belief, i.e., the subjective probability that an entity will
perform in a way likely to bring the expected benefit, or not do an unexpected harm.
Considering trust only as a subjective probability leaves out an extremely impor-
tant concept related to trust, that of risk. This fact has also been the catalyst of a vi-
gorous debate between economists and social psychologists [3]. In circumstances
where one entity relies on another, trust choices include a certain level of risk. Josang
[5] defines two different types of trust, Reliability and Decision trust. The former
covers the aspect of trust as stated by Definition 1a. The latter considers the risk
brought about by the uncertainty of transactional outcomes and is used to extend our
first definition:
Definition 1b. Trust is the extent to which one entity is willing to depend on oth-
ers‟ decisions and actions, accepting the risk of a negative (undesired) outcome.
2.2 Online representatives of trust
Despite the relatively interchangeable use of trust and reputation in the research
community, it is essential to understand the difference between the two.
Definition 2a. Reputation is the general standing of the community about an enti-
ty‟s trustworthiness, based on the past behavior, performance, or quality of service of
that entity, in a specific context, i.e., a domain of interest.
Definition 2b. A system that facilitates the process of calculating and managing
reputation is called a reputation system.
Hence, reputation is the amount of context-aware trust, i.e., a quantitative represen-
tation of the trust that the society places in an entity, bound by the domain of interest.
In addition to reputation systems, we consider rating and recommendation systems to
also be online representatives of trust. We define them as follows:
Definition 3. Rating systems manage the evaluation or assessment of something, in
terms of quality, quantity, or a combination of both.
Definition 4. Recommender systems are a subclass of information filtering sys-
tems that seek to predict the rating or preference that a user would give to an item or
a social element they had not yet considered, using a model built from the features of
an item (content-based) or the user's social environment (collaborative filtering) [6].
We use the terms trust mechanisms and RRR systems interchangeably in this work.
3 Related work
3.1 Social
Some of the work done in social and behavioral sciences that inspired computational
trust research was discussed in the previous section [1–3]. Neuroscience has also re-
vealed that emotions and cognition, present in different areas of the human brain,
interfere with each other in decision making, often resulting in a primacy of emo-
tions over reasoning [7]. A very similar, although deceptively simple idea stands be-
hind the outstanding work in Perceptual Control Theory: our perceptions are the only
reality we can know, and the purpose of all our actions is to control the state of this
perceived world [8]. The psychology of making trust-related choices is directly re-
lated to how people think and feel, perceive and decide. The brain has developed
complex mechanisms for dealing with threats and risks. Understanding how it works
and when it fails is critical to understanding the causal loop between trust-related
perceptions and trust-related choices. An area with remarkable results about the irra-
tionality, bias, and unpredictability of human actions in various circumstances and
mindsets is Behavioral Economy [9–11]. In the context of preferential reasoning, their
analyses show that users are often unaware of their taste, even for experiences from
previously felt outcomes. Not only this reveals that taste is much more subtle than
preference, but it shows that preference itself is not a stable property of human rea-
soning [12]. Experiments on persistency of user preferences about identical items at
different instances of time proved significant fluctuation in the repeated preferential
choices [10][13]. In contract and utility theory, the potential of employing trust me-
chanisms for dealing with information asymmetry was recognized long ago. When the
possibility of post-contractual opportunism creates a context of moral hazard, trust
mechanisms are employed for sanctioning undesired behavior. Another context of
information asymmetry is adverse selection, and arises when one is required to choose
a transaction partner whose type (good or bad) is unknown. In his work, Akerlof ana-
lyzes the effect of social and trading reputation on transaction outcome and market
maintenance [14]. The study demonstrates that goods with low quality can squeeze
out those of high quality because of the information asymmetry present in the buyers‟
and sellers‟ knowledge about the products – the problem of so called “lemon mar-
kets". Reputation mechanisms would balance this asymmetry, helping buyers make
better-informed decisions by signaling the behavioral types of sellers, and at the same
time providing incentives for sellers to exchange high-quality goods. Thus, Akerlof
makes an instructive distinction between the signaling and the sanctioning role of
RMs, which was only recently considered in computer science.
3.2 Technological
Understanding the behavioral implications of users in the field of computational trust
is crucial, as the user factor in the processes met in online trust mechanisms is omni-
present. As online representatives of trust, RRR systems are also assigned the role of
“devices” that help decision-making under information asymmetry (reputation and
rating systems), and information overload (recommendation systems). In the former
case, they have sanctioning and signaling role, and in the later case – directing and
filtering role. The goal of employing RRR systems is to reduce the complexity that
arises from information overload, and to lower the uncertainty present in the contexts
of information asymmetry. Depending on the general context, however, the combina-
tion of the three (reputation, recommendation, and rating) requires different structur-
ing to achieve the desired goals. This is discussed further in the following section.
RRR systems rely to a great extent on preference inputs from users. Bias in these
inputs may have a cascading error effect on the performance of the applied algo-
rithms. This does not only affect the accuracy of the results, but it also influences the
perceived system reliability. Hence, user preferences are malleable and affect system
performance, but they are also largely influenced by the information provided by
RRR systems. Yet, biased behavior, its causes and effects, are relatively unexplored
issues in the field. But the fact that only a narrow set of cognitive biases has been
tackled by the research community does not imply there are no significant studies
made in this regard. In [18] and [19], authors investigate the so called self-selection
bias, whereby users only rate the items (movies) they like most, causing extremely
high average rating for the rated items. Such ratings are representative only for a spe-
cific group of users, but do not truthfully depicture the item‟s general quality. Fur-
thermore, self-selection bias was proved to not only be a transient phenomenon, but
also a steady state in the system [15]. Although seemingly absurd, there is also one
positive implication from this result of “self-selection bias” sustainability: if there was
a strategy in the managerial principles of some company to cause inflated ratings for
certain products by the “first-mover effect” [17], this effect would be flattened out on
the long run. Such issues may appear to have purely economic nature, but they se-
riously compromise the reliability and performance of current RRR systems. The high
impact of online reviews on product sales was demonstrated in [18], uncovering some
of the motifs behind companies‟ efforts to appear competitive on the market. Positive
rating bias was noted throughout systems of different nature. eBay is claimed to owe
its success to its reputation/feedback system, yet out of the 57% of the users that de-
cide to leave feedback, 99% issue positive feedback [19]. Moreover, large amount
(41%) of users prefers to stay silent rather than to leave a negative feedback. A pro-
posal to interpret silence as part of the user feedback was proposed in [20].
4 Cognitive Bias and its loops of causality in RRR systems
Cognitive bias describes a replicable pattern in perceptual distortion, inaccurate
judgment, or illogical interpretation [10]. Clearly, bias can be noticed in both percep-
tions and actions, and the two are bound by the way people process their perceptions
in order to take an action (including non-action). The paradox that arises in RRR sys-
tems is that, although user preferences are overly biased and affect the system per-
formance and reliability, the preferences themselves are largely influenced by the
ratings and the results provided by RRR systems. In this section, we explore the
causes and effects related to cognitive bias in RRR systems, and we identify four
major factors that influence these causal loops of biased behavior.
4.1 Context and Bias
Context is the set of circumstances or facts that surround a particular event or situa-
tion. Here, we analyze three general contexts in which an online interaction can take
place: one with pure collaborative elements, one with collaborative and competitive
elements, and one with collaborative, competitive, and monetary elements. Clearly,
each subsequent context includes the elements of the previous, implying added sys-
tem complexity. Therefore, it is crucial to understand which elements can be ade-
quately combined in order to meet a system design-goal, without encumbering its
performance and flexibility to the extent of edging users out.
Once a reputation score becomes part of users‟ profiles, the users themselves be-
come identifiable by their reputation, as if it guards their „online brand‟. Hence, main-
taining a stable identity is crucial for the joined reputation value to make sense. On
the other hand, being equipped with an online identity as a synonym for one‟s reputa-
tion gives rise to some new dimensions in the context of impression management.
Once reputation is used as a signal for the user behavioral type, pure collaborative
context ceases to exist. The requirement for conducting a successful impression man-
agement adds a competitive component to the system, and makes the reputation expli-
citly recognized as part of a person‟s social capital. Coupling this with the presence of
bias implies that inflated (overly positive) reputation values in a system devaluates the
reputation itself, as if the presence of reputation value defeated its purpose. Such a
situation creates the need for additional incentives that will shift the reference for a
good behavior from impression management to another context. What is often done in
this regard is applying monetary elements to provide incentives for a desired beha-
vior, which brings its own issues. Mixing purely collaborative context with monetary
elements was already proved to throw a shadow on both the social intentions and the
opportunity for monetary gain [10]. In addition to these more subtle influences, there
are more detrimental effects that arise from inadequately combining such context-
elements. Including monetary elements drastically increases the complexity of the
system, and introduces inter-locking dependencies between the trust mechanisms and
the outside environment. From a systemic perspective, this implies that the boundaries
of the system are open to additional disturbances [21]. As a result, claiming predicta-
bility of the users‟ reputation scores, or moreover, of user and system behavior, dimi-
nishes, and failing to recognize this leads to system degradation, and eventually, sys-
tem failure. This is also the core idea behind the Tragedy of Commons [22].
The soundness of the matters elaborated above was also demonstrated in practice.
By announcing its Partner‟s Program1 in May 2007, YouTube explicitly offered its
highest rated users to earn revenue from advertisements placed next to their videos.
This instantly triggered series of events of users blaming each other for using auto-
mated programs to inflate their videos‟ ratings2. While the effect of these blames is
related to the tragedy of commons [22], the effect of using programmed agents to
inflate one‟s own rating is known as the Cobra effect [23]: the solution of a problem
makes the problem worse. These effects are often result of systemic ignorance, and
only retrospectively analyzed by many system designers.
Pure collaborative contexts exist when a reputation is used internally in the system
(for e.g., to provide a reference or serve as a regulator for the flow of some system
processes), also implying keeping users‟ reputation private, or if the acquiring of rep-
utation is not bound to one‟s performance, i.e. it is not used for signaling purposes.
Clearly, when making the decision of which contextual elements to choose as part
of a system, context is intimately related to risk. With the addition of each element,
the complexity increases, and the perceived risk is hardened by additional factors.
Furthermore, privacy appears worthy of consideration as an option for limiting the
1 http://www.youtube.com/creators/ 2 http://gigaom.com/video/real-or-robot-the-lisanova-controversy/
detrimental effects of added complexity. The goal of this study is to have a holistic
look on RRR systems through the defined factors, rather than analyzing each of them
independently. Therefore, the next section examines the link between risk and bias.
4.2 Risk and Bias
Risk is conceived as the possibility of triggering unexpected, unlikely, and detrimen-
tal consequences by means of a decision attributable to a decision maker [21]. Uncer-
tainty is part of every online interaction. The extent of uncertainty, the expected utili-
ty, and the cost required for performing an action, influence the perceived risk a
transaction brings. The field of prospect theory offers an incredible amount of expe-
rimental work demonstrating the myriad of cognitive biases that people exhibit when
faced with risk and uncertainty [11]. A phenomenon that binds risk and uncertainty is
the so called pseudo-certainty effect, which reflects the tendency of people to be risk-
averse if the expected outcome is positive, but risk-seeking when they want to avoid
negative outcomes. While it is a curiosity to inspect the properties of each bias inde-
pendently, in reality biases are often coupled together, acting as both the cause and
the effect of human perceptions and actions. The property of non-linearity we as-
signed to trust systems implies that complementing few biases together does not mean
that their causes and effects will work in an additive fashion. Prospect theory has
demonstrated that people underutilize consensus information, and when given de-
scriptive facts about the quality of a person, they make choices regardless of the sta-
tistics offered about that person [11]. Josang et al. provide a formal proof of this phe-
nomenon known as the base-rate fallacy, and give a formal framework for accounting
for this fallacy in a computational setting [24]. Information offered in RRR systems is
to a great extent a statistic produced by the collective efforts of the community mem-
bers. This information can be represented in various ways - numerically, descriptive-
ly, as a single or multi-valued component, or as a combination of those. Therefore, it
is important to explore how users perceive different types of information, and whether
the descriptive and the numerical representation of an entity‟s quality can collide in
users‟ perception. This is something we also investigate in our experimental study.
4.3 Dynamics and Bias
To some extent, we already touched on the issue of dynamics and bias when discuss-
ing the inconsistency of user preferences over time [10]. Here, we stay more in the
context of trust and reputation, and connect the dynamics-factor with the previously
defined – risk and context.
From the definitions of trust and reputation (Definitions 1and 2), it becomes clear
that the dynamics of trust differs from the dynamics of reputation. This discrepancy
cannot be captured by any model, as both trust and reputation are in reality intangible
matters. Yet, the social models and protocols for detecting malice seem to be success-
ful. In RRR systems, one way the dynamics of reputation is embedded in the models
is through discounting the relevance of gathered information by some time-factor.
However, such approach disregards the importance that some information had in the
past in terms of its impact on the outcome. In other words, discounting by recency and
frequency is not equal to discounting by impact. Doing the former may provide disin-
centives for the users who take important actions at a lower rate. Closely related to
this issue is the bias of rosy retrospection, which is a tendency to rate past events
more positively than they were actually rated when the event occurred [11]. This re-
flects the importance of accounting for the time-interval between item-consumption
and provided feedback about that item. Unfortunately, our current experimental study
does not tackle any of the concepts related to dynamics. However, as a disintegrable
part of the bigger picture of online trust mechanisms, dynamics must be taken into
consideration. Our future work will devote more attention to this factor.
4.4 Privacy and Bias
In [25], authors show that individual interpretations of trust and friendship vary, and
the two concepts are correlated to other characteristics of a social tie and to each other
in a non-symmetric way. Furthermore, they provide evidence that raters consider how
a ratee‟s reputation might be affected by the feedback. This fear of bad-mouthing and
reciprocation in the context of impression management is directly related to the fear
of retaliation in e-commerce systems [19]. Together, these factors also help to explain
patterns like higher reciprocity in public ratings, and the near absence of negative
ratings. Given this reluctance to publicly leave negative feedback, the question arises:
why offer multi-valued choice for item evaluation in the first place. Moreover, why
showing it publicly if it affects the users‟ decisions to an extent that makes it useless.
Closely related to privacy and the context of impression management are the con-
cepts of individual and group behavior, and similarly, individual and group bias.
Whereas most of the biases we mentioned so far were characteristic for an individual,
we would fall extremely short on a useful discussion if we do not touch upon group
behavior as well. After all, building reputation is essentially a social process, regard-
less of the fact that trust individuals cherish for one another underlies this process.
In [26], authors study the anchoring effect that item ratings have on user prefe-
rences. They find that users‟ inclination towards providing positive feedback is addi-
tionally amplified if users see the current rating that an item got by the rest of the
community members. That people imitate, or do what others do, especially when
having no determined preferences, is nothing new. This is often the cause of what is
known as group polarization, bandwagon effect, or herd behavior, depending on the
field of study that identifies it [9]. In the context of RRR systems, a study on group
polarization on Twitter showed that, like-minded individuals strengthen group identi-
ty [27]. In other words, when part of group situations, people make decisions and
form opinions that are extreme than when they are in individual situations. For RRR
systems, this implies that it is not only important to acquire a significant amount of
user feedback, but to also investigate whether this amount of user opinions was in-
ferred from sufficient number of independent sources. A formal apparatus for resolv-
ing such issues in a computational setting can be found in [28], where the author pro-
vides a framework for reasoning with competing hypothesis using Subjective Logic.
The following section describes the experimental study and analyzes the findings
through the factors defined above. Moreover, it proposes some ways to address the
revealed issues.
5 Experimental work
Objectives: The major objective of this survey was to provide data that will help us
investigate the compatibility between users‟ perceptions of the RRR system and its
design objectives, but also reveal new directions for reasoning about the inter-
relations between users and systems. The main questions we aim to answer are:
─ Which descriptive model resembles closest user perception of numerical feedback?
─ How does a slight difference in the „tone‟ of presented choices in the two descrip-
tive models influence users‟ decision?
─ How does the presence/absence of different contextual elements influence user
choice, and is it related to the nature (numerical or descriptive) of the alternatives?
Design and Methodology: Two types of methods were used to gather the necessary
data – an online survey method3 was chosen for better geographical spread of respon-
dents, speed of data collection and independence of participants‟ opinions; and direct
(one-on-one) interviews were performed to capture some of the subtleties that are
ungraspable by only observing the outcomes of the systems. Such subtleties are the
discrepancy between preferences of the majority of individuals and the group prefe-
rence, the huge difference in the choices made as new contextual elements are intro-
duced or taken away, the difference and inter-dependence between trustworthiness
and acquaintanceship, etc. Each survey contained the same two questions, but offered
slightly different evaluation choices. The questions represented real reviews for HP
Laptop taken from Epinions4. To not disturb the flow of the paper, they are given in
the Appendix. All three groups were asked to rate the reviews for their usefulness.
The first group (Survey 1) was asked to give a numerical rating on a scale 1 to 5 (1 =
lowest and 5 = highest rating). The second and the third group were offered descrip-
tive evaluation choices. However, among the possible answering choices for Survey
2, there were also such that stated explicit negative experience (Not useful at all),
whereas the answering choices for Survey 3 varied from neutral to positive.
Table 1. Statistics about the experimental setting
Respondents: The experimental work is conducted over a population (both female
and male) of 86 people, at the age of 20 – 50. Its completion required no special tech-
nical knowledge, and subjects had no difficulty understanding the assignment. Res-
pondents were divided into 5 groups. Three groups of 22 people were formed for each
survey type. Table 1 summarizes these statistics. 30 of the 66 survey respondents
3 www.surveymonkey.com 4 www.epinions.com
Survey Info.
Respond.
Info.
Type of Survey
Survey 1 Survey 2 Survey 3
1-5 Num. Feedback Neg.Descriptive Neutral-Pos. Descriptive Total
Answering choices
1
2 3
4 5
Not useful at all
Hardly useful Somewhat useful
Quite useful Extremely useful
Neutral
Somewhat useful Quite useful
Very useful Extremely useful
Responded 22 22 22 66
were also additionally interviewed. The results from these interviews are presented in
the final subsection. In addition to the three groups, another 20 users were asked to
independently evaluate each review (10 users per review). They expressed their eval-
uations both numerically, as in Survey 1, and descriptively, as in Survey 2.
Results: The following subsections show the major findings from this experimental
study. Although the experiment was of relatively small scale, some interesting results
were revealed with respect to the given objectives.
5.1 Distinction bias
As a first step in revealing the potential presence of cognitive bias in our experimental
setting, we compare the results from the individually evaluated reviews with those
obtained by the three surveys, where the two reviews were put together for evaluation.
The goal is to reveal the potential presence of a distinction bias, manifested as a ten-
dency to view two options as more dissimilar when evaluating them simultaneously
than when evaluating them separately. This bias is often exploited in commerce sce-
narios, when sellers aim to sell a certain product (anchor) by placing it along with
another - decoy product that appears as the worse option when put together with the
anchor [10]. Fig.1 presents the results from the 20 independent evaluations of the two
reviews. The horizontal axis shows the rating values, whereas the vertical – the num-
ber of users who provided the rating. As shown, both reviews were evaluated as
equally useful, 2 on a 1 to 5 scale, and descriptively qualified as Somewhat Useful.
a) b)
Fig. 1. Rating distribution for each review by offline users a) Numerically b) Descriptively
Fig.2 shows the difference in ratings between the two reviews for each of the surveys.
Clearly, that they do not match with those shown in Fig1. In all three cases, Review 1
was evaluated as being better than Review 2. The mean values for the ratings of Re-
view 1 and Review 2 for the three surveys are given in Table 2. Since preferences are
often formed through distinction between given options, joint evaluation of recom-
mendations may often result in a choice mismatch. The consequence is that the choice
that appeared as the best option may not provide the best user experience, leading to
dissatisfaction. Clearly, the issues of distinction bias and anchoring effect are impor-
tant to account for in the design of RRR systems. The question is: what is the cause of
those biases, what their effects are, and how to account for them in practice. The next
section aims to provide the answer.
0
2
4
6
8
1 2 3 4 5
0
1
2
3
4
5
6
7
1 2 3 4 5
Table 2. Mean value for the ratings of the surveys
1-5 Num. Feedback Neg. Descriptive Neutral/Pos. Descriptive
Review 1 3,727 3,545 3,636
Review 2 2,5 3,18 1,95
a) b) c)
Fig. 2. Rating Distribution for the two reviews for a) Survey 1, b) Survey 2 and c) Survey 3
5.2 Numerical-Descriptive discrepancy and Positive Bias
In order to address distinction bias and anchoring effect, we must first understand if
and how the presented evaluation choices affect the user opinion. This section inves-
tigates which of the descriptive surveys comes closest to the one with numerical rat-
ings. The practical implication is in exploring if the current RRR systems that offer
numerical ratings really match the user understanding about the meaning of those
ratings. For that purpose, we compare the rating distributions of the two reviews for
Survey 1 to those of Survey 2 and Survey 3.
a) b)
Fig. 3. Difference in Rating Distribution between Survey 1 and 2 for a) Review1 b) Review2
The results are shown in Fig.3 and Fig.4 respectively. It can be noticed that there is a
good match for the distributions of Review 1 in both of the cases. However, compared
to the distributions on Fig.1, there is still a great difference between the average rating
value for Review 1 (1.9) provided by the independent user evaluations, and the aver-
age rating value for Review 1 in Survey 1 (3.727), Survey 2 (3.545) and Survey 3
(3.636). We identify the following causal link: distinction bias is caused when Review
1 and Review 2 are put together for evaluation, whereby Review 1 appears as the
0
2
4
6
8
10
12
1 2 3 4 5
0
2
4
6
8
10
12
1 2 3 4 5
0
2
4
6
8
10
1 2 3 4 5
0
2
4
6
8
10
12
1 2 3 4 5
0
2
4
6
8
10
12
1 2 3 4 5
better option; moreover, Review 1 also appears as one of exceptional quality, leading
to a positive bias and exaggerated positive rating.
b) b)
Fig. 4. Difference in Rating Distribution between Survey 1 and 3 for a) Review1 b) Review2
5.3 Positive bias and the Framing effect
In this section, we analyze the two surveys with descriptive evaluation choices. They
differ slightly by the tone of positivism, although both offer five choices. Survey 2
contains more explicit negative statements, while the choices in Survey 3 vary from
neutral to positive. The question we want to answer is How does a difference in de-
scriptive choices affect users evaluation? To do that, we compare the rating distribu-
tion of the reviews for Survey 2 and Survey 3. The results are given in Fig.5a) and b).
a) b)
Fig. 5. Difference in Rating Distribution between Survey 2 and 3 for a) Review 1; b) Review 2
They reveal large difference in the evaluations of the reviews between the surveys,
although the offered choices differ only slightly. This demonstrates that people tend to
draw different conclusions from the same information, depending on how that infor-
mation is presented, known as the framing effect. On Fig.5a), this effect is demon-
strated as a slight smoothing of the exaggerated positives. Complementing these re-
sults with those shown on Fig.2b), we can conclude that for Survey 2, the ratings for
the two reviews also come close to each other. One interesting result is the high mean
rating value for Review 2 in Survey 2. Compared to the results on Fig.1, we see that
the presence of a positive bias is higher when users are offered to chose between neg-
0
2
4
6
8
10
1 2 3 4 5
0
2
4
6
8
10
12
1 2 3 4 5
0
2
4
6
8
10
12
1 2 3 4 5
0
2
4
6
8
10
1 2 3 4 5
ative-positive evaluations, compared to a setting where they are required to choose
between neutral-positive. In the next section, we use these findings to form our pro-
posal on how to address some of the issues of the explored biased behavior.
5.4 Proposal: Hidden signals and “shades of grades”
The remarkably higher number of neutral evaluations demonstrated by the experiment
(Fig.5b)), which is even higher than the positive evaluations for Review 2 (Fig.5a)),
not only confirms the reluctance of users to give negative ratings, but also points out
the importance of neutral vote as a connector between negative and positive. In addi-
tion to introducing neutral as a “shade” between negative and slightly positive, anoth-
er finding in our study is that introducing Very Useful as a shade between Quite Use-
ful and Extremely Useful shifts the exaggerated positives towards lower ratings. Both
of these effects smooth the effects from positive bias, but also better capture people‟s
perception of the offered choices. This might also explain why the offline model of
social reputation succeeds to detect untrustworthy individuals without requiring con-
sensus on someone‟s trustworthiness. The offline world offers numerous opportuni-
ties to pick on the hidden clues behind people‟s intentions. We consider the idea of
introducing such hidden signals in the online RRR systems worth exploring.
The framework we propose for addressing the presented issues of biased behavior
can be summarized as follows: first, by accounting for the specific context elements,
we propose disentangling the collaborative, competitive and the monetary elements
when deciding on the RRR design. Second, by accounting for those context elements,
a decision should be made about which of them are desirable as public features. Then,
it is crucial to determine the right representation of publicly displayed features in a
way that fits the users‟ perception of the feature‟s meaning. Finally, by introducing
hidden signals and shades of grades about the qualitative types of the entities, a better
and more truthful distribution of the results can be obtained.
5.5 Closing the loop: The market of lemons on the market of opinions?
This section will close the loop of our study on cognitive bias by returning to the
point where we started the discussion – context. The impact of context in the forma-
tion of cognitive bias required more interactive work with the respondents of the sur-
vey. Therefore, we additionally interviewed 30 people. The results are the following:
Q1. If you know that, depending on the ratings given for their reviews, reviewers will
get proportionally higher/lower amount of money, would you give the same grade?
22 answered: No, one or two grades lower. 1 was Not Sure; 7 answered Yes;
Q2. What if you yourself were a reviewer and the amount of money you would get
depended on the amount of money other reviewers for the same product get?
27 answered Definitely a lower grade; 3 answered Still the same grade.
The 3 who gave the 2b) answer were additionally asked:
Q3. If you see that the opinion you consider of low quality is the one that got the
highest number of votes, would you reason the same the next time?
2 answered: No; 1 answered Yes.
The purpose of these questions was to investigate the reasoning of the respondents
as they were required to switch between contexts. The interview is of very small
scale, but still pointed to new directions for reasoning about user behavior through the
elaborated contextual elements. There is, however, a deeper meaning of the obtained
results: the exaggerated positives, the slight change in evaluation options followed by
high difference in evaluations, and the rest of the biases we explored, may be merely
the effects of user behavior in RRR systems. Different combination of particular con-
textual elements leads to different manifestation of the effects from those biases. In-
troducing monetary elements must be done with great caution, as it may cause infor-
mation to be treated as a limited resource of monetary value, or as a trading resource
in the process of acquiring social capital. This in turn leads to squeezing entities
(items, agents, users) of potentially high quality out of the system, leading to what we
refer here to as “the market of lemons on the market of opinions”.
6 Concluding Remarks and the Way Forward
Trust is a feeling, a model, and a reality, with people in their core. Understanding how
they work and how closely they resemble each other is essential for their design prac-
tices. Our work detected and analyzed context, risk, dynamics and privacy as factors
that influence both the work of RRR systems and the users‟ understanding of that
work. We explored the relation between RRR systems and user behavior with respect
to those factors and analyzed few types of biases exhibited through the users‟ online
experiences with RRR systems. In an experimental study that included 86 users, we
found that users exhibit distinction bias, positive bias, anchoring effect, and framing
effect. These have not been investigated in such a holistic manner for any of the cur-
rent RRR systems. Based on the identified factors and the findings of the experimen-
tal study, we proposed a framework for tackling some of the issues attributed to users‟
biased behavior and address the possibility of employing hidden signals and shades of
grades in the RRR systems for the purpose of capturing some of the detected biases.
Our future work will concentrate on investigating more the factor of dynamics. As
we already referenced Perceptual control theory and Subjective Logic as an apparatus
that provides formal reasoning with respect to human beliefs and perceptual behavior,
the major part of our work will be directed towards joining the ideas of the two and
employing them to formalize trust relationships under biased user behavior.
Acknowledgments. Authors wish to thank Tomaž Klobučar, Dušan Gabrijelčič and
Borka Jerman Blažič for the devoted time and energy, their opinions, critiques, and
immensely beneficial discussions regarding the topics of this work.
References
1. R. B. Zajonc, „Feeling and thinking: Preferences need no inferences‟, American Psycholo-
gist, vol. 35, no. 2, pp. 151-175, Feb. 1980.
2. John Conlisk, „Why Bounded Rationality?‟, Journal of Economic Literature, vol. 34, no.
2, pp. 669-700, 1996.
3. C. Castelfranchi and R. Falcone, „Trust is much more than subjective probability: Mental
components and sources of trust‟, 32nd Hawaii International Conference on System
Sciences - Mini-Track on Software Agents, Maui, vol. 6, 2000.
4. D. Gambetta,„Can We Trust Trust?‟Trust:Making and Breaking Cooperative Relations
5. A. Josang, R. Ismail, and C. Boyd, „A survey of trust and reputation systems for online
service provision‟, Decision Support Systems, vol. 43, no. 2, pp. 618-644, Mar. 2007.
6. F. Ricci, L. Rokach, and B. Shapira, „Introduction to Recommender Systems Handbook‟,
in Recommender Systems Handbook, F. Ricci, L. Rokach, B. Shapira, and P. B. Kantor,
Eds. Boston, MA: Springer US, 2011, pp. 1-35.
7. R. MacMullen, Feelings in history, ancient and modern. Regina Books, 2003.
8. W. T. Powers, Behavior: The Control Of Perception,2Benchmark Publications, Inc., 2005.
9. D. Kahneman, „Maps of Bounded Rationality: Psychology for Behavioral Economics‟,
American Economic Review, vol. 93, no. 5, pp. 1449-1475, 2003.
10. D. Ariely, Predictably Irrational: The Hidden Forces That Shape Our Decisions, 1st ed.
HarperCollins, 2008.
11. D. Kahneman and A. Tversky, „Subjective probability: A judgment of representativeness‟,
Cognitive Psychology, vol. 3, no. 3, pp. 430-454, Jul. 1972.
12. D. Ariely, G. Loewenstein, and D. Prelec, „Tom Sawyer and the construction of value‟,
Journal of Economic Behavior & Organization, vol. 60, no. 1, pp. 1-10, May 2006.
13. D. Cosley, S. K. Lam, I. Albert, J. A. Konstan, and J. Riedl, „Is seeing believing?: how re-
commender system interfaces affect users‟ opinions‟, in Proceedings of the SIGCHI confe-
rence on Human factors in computing systems, New York, NY, USA, 2003, pp. 585–592.
14. G. A. Akerlof, „The Market for “Lemons”: Quality Uncertainty and the Market Mechan-
ism‟, The Quarterly Journal of Economics, vol. 84, no. 3, pp. 488-500, 1970.
15. M. Kramer, „Self-Selection Bias in Reputation Systems‟, in Trust Management, vol. 238,
Springer Boston, 2007, pp. 255-268.
16. X. Li and L. M. Hitt, „Self-Selection and Information Role of Online Product Reviews‟,
Information Systems Research, vol. 19, no. 4, pp. 456 -474, Dec. 2008.
17. M. B. Lieberman and D. B. Montgomery, „First‐mover advantages‟, Strategic Man-
agement Journal, vol. 9, no. S1, pp. 41-58, Jun. 1988.
18. J. A. Chevalier and D. Mayzlin, „The Effect of Word of Mouth on Sales: Online Book Re-
views‟, Journal of Marketing Research, vol. 43, no. 3, pp. 345-354, 2006.
19. P. Resnick and R. Zeckhauser, „Trust among strangers in internet transactions: Empirical
analysis of eBay‟ s reputation system‟, in Advances in Applied Microeconomics, vol. 11,
Bingley: Emerald (MCB UP ), pp. 127-157.
20. C. Dellarocas and C. A. Wood, „The Sound of Silence in Online Feedback: Estimating
Trading Risks in the Presence of Reporting Bias‟, Management Science, vol. 54, no. 3, pp.
460 -476, Mar. 2008.
21. N. Luhmann, Risk: a sociological theory. Transaction Publishers, 2005.
22. G.Hardin, Tragedy of the Commons,Science,vol.162,no.3859,pp.1243-1248,Dec. 1968.
23. H.Siebert,DerKobra-Effekt :wie man Irrwege der Wirtschaftspolitik vermeidet:Piper, 2003.
24. Audun Josang and S. O‟Hara, „The base rate fallacy in belief reasoning‟, in 2010 13th
Conference on Information Fusion (FUSION), 2010, pp. 1-8.
25. L. Adamic, D. Lauterbach, C.-Y. Teng, and M. Ackerman, „Rating Friends Without Mak-
ing Enemies‟, Proceedings of the Fifth International AAAI Conference on Weblogs and
Social Media, pp. 1-8, 2011.
26. J. Zhang, „Anchoring effects of recommender systems‟, in Proceedings of the fifth ACM
conference on Recommender systems, New York, NY, USA, 2011, pp. 375–378.
27. S. Yardi and D. Boyd, „Dynamic Debates: An Analysis of Group Polarization Over Time
on Twitter‟, Bulletin of Science, Technology & Society, vol.30,no.5,pp.316 -327,Oct. 2010.
28. S. Pope, „Analysis of Competing Hypotheses using Subjective Logic‟, Systems Technolo-
gy, no. June, pp. 13-16, 2005.
Appendix: Survey Questions (Reviews)
Review 1:
User Rating: OK; Ease of Use: 2;Quality of Tech Support: 1
Pros:1)Intel core i3 processor;2)Finger Print Sensor;3)Battery life
Cons:1)Build material;2)Intel integrated graphics;3)320GB hard disk space
The Bottom Line: This laptop is not recommended to any one, because of poor support quality.
The products of HP were good 3 years ago. People loved them for their looks, reliability, per-
formance and budgeted price. But after 2008 the reliability ratio of HP as compared to its com-
petitors is far below in almost all aspects including price, performance, looks and the most
important reliability. I myself was a big fan of this company but after facing the failure of con-
secutively two the PCs' from this company, I changed my view. My wife had this laptop and I
bought this laptop for her as her birthday present three months ago, but just after two months,
the screen got dead spot (small black spots) and the hing connecting the screen with keyboard
also got broken. Still the laptop was running fine but last week its keyboard also stopped work-
ing, which sucks. All the money I paid for it gone in vain. My experience about: PROCESSOR:
Intel core i3 processor clocked at 2.1 GHZ works really very fine for multitasking but it is not
designed to handle more tasks. This pc is fine for web-surfing, word-processing, office work,
watching movies. GRAPHICS: Intel integrated graphics are not capable of running blue ray
movies silky smooth. BUILD MATERIAL: This laptop is made of seriously very cheap plastic.
Its glossy surface is just a fingerprint magnet and you have to clean up the laptop after every
single use. The glossy screen also causes panic while watching movies in sunlight. LACK OF
HDMI PORT: This laptop lacks some of the most important port which is included in the lap-
tops of this price range, i-e; HDMI port. The sound quality is good, battery life is also impres-
sive and last about 3 hours (6 cell) even when watching movies. The security feature like Fin-
ger Print Sensor works great. Its light weight makes it easier to carry this laptop everywhere but
it can't be called ultraportable laptop. Over all this laptop isn't recommended to any one wheth-
er the person is student or house wife.
[Recommended: No]
Review 2:
User Rating: Excellent; Ease of Use: 5; Quality of Tech Support: 4
Pros: I would buy another computer from HP
Cons: multiple hardware faults. (i.e hard to type and sensitive mouse pad)
HP ProBook 4530s is a extremely good product. The computer is lightweight and rugged. It's
able to be put in a backpack and stuffed wherever you need to go. When i bought and received
the computer i liked how the computer didn‟t come with a lot of extra software junk. Next, i
like having a numeric pad and i will continue to buy computers with them from now on. For
just web browsing and playing light computer games this computer is extremely fast. The key
pad is not sensitive enough and i have to hit the keys hard a lot. Also if you have multiple fin-
gers close to the mouse pad, the mouse pad goes crazy due to its many settings. The speaker
system is excellent the loudest system i have heard on any computer system. The software for
the finger ID gets weird every now and then i have to input my fingers a couple of times to log
in. The settings are easy to change on the computer. The battery life is average to other com-
puters. The computer hardware is nicely laid out and easy to find. Computer is worth the price i
paid for it.
[Recommended: Yes]